[WANDY666]

At high output rates, PD master spends CPU constructing per-token API models and forwarding already-ready output. Serialize common chat/completion deltas directly, group chat SSE events at ready-batch boundaries, and merge ready completion text per choice before serialization.

Single-choice PD requests consume the source generator directly instead of using a forwarding task and queue, with source cleanup on cancellation. WebSocket receives keep their timeout without creating a task per message, tool metadata is computed once per stream, and length validation accepts a token count directly.

The first completion chunk, batch end, finish, or 32-token/64K-character limit flushes immediately; there is no timer waiting for another model token. Preserve chat event ordering, per-choice output text, tool arguments, echo/logprobs payloads and token-based usage. Flush buffered output before stream errors. Merge current main while retaining its prefill/cache metadata and finish-status handling.

Validation: 379 tests passed across streaming contracts, completion batching, PD continuation/cache/lifecycle, multi-choice cancellation, application stop, tool calls and reasoning. Black 21.12b0, Flake8 6.1.0 and diff whitespace checks passed for all 9 changed files.

Historical performance evidence for the added completion batching + single-choice bypass is recorded in speedup_1012.md: in the controlled 1P1D H100 experiment (16K input, 4096 output, concurrency 512), output TPS increased 26,730.2 → 30,066.7 (+12.48%) and mean request TPOT fell 15.4485 → 13.0199 ms (−15.72%). Each side was one run; this measures those two changes on that deployment, not the whole PR on current main. No new service benchmark was run for this update.